Takens Bogdanov bifurcation on the hexagonal lattice for double-layer convection

Physica D 129 (1999) 171 202 Takens Bogdanov bifurcation on the hexagonal lattice for double-layer convection Yuriko Yamamuro Renardy a,, Michael Renardy b, Kaoru Fujimura c,1 a Department of Mathematics and Interdisciplinary Center for Mathematics, Virginia Polytechnic Institute and State University, Blacksburg, VA 24061-0123, USA b Department of Mathematics and ICAM, V.P.I & S.U., Blacksburg, VA 24061-0123, USA c Japan Atomic Energy Research Institute, Tokai-mura, Ibaraki 319-11, Japan Received 20 April 1998; received in revised form 10 December 1998; accepted 7 January 1999 Communicated by F. H. Busse Abstract In the Bénard problem for two-fluid layers, Takens Bogdanov bifurcations can arise when the stability thresholds for both layers are close to each other. In this paper, we provide an analysis of bifurcating solutions near such a Takens Bogdanov point, under the assumption that solutions are doubly periodic with respect to a hexagonal lattice. Our analysis focusses on periodic solutions, secondary bifurcations from steady to periodic solutions and heteroclinic solutions arising as limits of periodic solutions. We compute the coefficients of the amplitude equations for a number of physical situations. Numerical integration of the amplitude equations reveals quasiperiodic and chaotic regimes, in addition to parameter regions where steady or periodic solutions are observed. 1999 Elsevier Science B.V. All rights reserved. 1. Introduction The two-layer Bénard problem concerns instabilities that arise when a two-layer system is heated from below. An instability may take place when the temperature difference between the upper and lower walls reaches a threshold value, or it may take place even at very low Rayleigh numbers due to the stratification in the fluid properties. A time-periodic onset of instability is possible, and there are two very distinct mechanisms involved. Each mechanism consists of two modes competing with each other to set up the oscillations. The first is associated with a deformable interface, and the oscillations are due to the competition between the Bénard instability (due to a bulk mode destabilization) and a stabilizing interface (due to an appropriate stratification in fluid properties). The eigenmodes consist of time-periodic convection cells which extend through both fluids. This mechanism was studied in [1,2]. Corresponding author. Tel.: +1-540-231-8258; fax: +1-540-231-5960; e-mail: renardyy@math.vt.edu. 1 Permanent address. Department of Applied Mathematics and Physics, Tottori University, Tottori 680, Japan. 0167-2789/99/$ see front matter 1999 Elsevier Science B.V. All rights reserved. PII: S0167-2789(99)00007-X

172 Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 Fig. 1. Sketch of cross-over of eigenvalues for a degenerate double zero eigenvalue over a range of bifurcation parameters, R and l 1 to be defined later as the Rayleigh number and the dimensionless depth of the lower liquid. Fig. 2. Sketch showing cross-over of real eigenvalues to form complex conjugates at two Takens Bogdanov points. A second mechanism for oscillatory onset arises when interface deformation is negligible. In this case, the critical eigenmode consists of two cells arranged one above the other, that is, one in each fluid [3 6]. This mechanism has recently drawn the interest from experimental physicists [7,8,24]. Oscillatory onset can occur if the effective Rayleigh numbers for each layer are roughly equal. The fact that a criterion of nearly equal Rayleigh numbers applies suggests that the occurrence of complex eigenvalues can be thought of as resulting from the near crossover of two real eigenvalues. Indeed, there are special values for the ratios of fluid properties for which the problem is self-adjoint, so real eigenvalues would then cross without becoming complex. This observation has led to studies of bifurcations in the neighborhood of a double zero eigenvalue which represents such a crossover [9,10]. Fig. 1 is a sketch of the situation for the two-layer problem, with the vertical axis representing a parameter such as the Rayleigh number, and the horizontal axis representing another parameter, such as the dimensionless depth of the lower liquid. In this paper, we focus on a less degenerate situation, which is that of a Takens Bogdanov point. Fig. 2 is a sketch of this in terms of the relevant parameters for the two-layer problem, analogous to Fig. 1. Such a point represents the merger of two real eigenvalues which then form a complex conjugate pair. This is less degenerate than the crossover of two real eigenvalues, which can be thought of as arising from the merger of two Takens Bogdanov points. Since the window for the complex eigenvalues in Fig. 2 is small, this picture may be thought of as a perturbation of the situation in Fig. 1. Takens Bogdanov bifurcations without symmetries or with simple symmetries such as Z 2 or O(2) are wellstudied and much is known about periodic and homoclinic solutions [11 14]. However, the case of the hexagonal lattice, which is of interest for the Bénard problem, remains to be investigated. Below, we discuss how to derive the amplitude equations governing the neighborhood of a Takens Bogdanov point, and how to obtain them by a limiting procedure when the frequency of a Hopf bifurcation tends to zero. We then analyze the existence of bifurcating solutions of various symmetry classes. For the case of Hopf bifurcation,

Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 173 Fig. 3. Sketch of the region of validity of the analysis carried out in this paper in ɛ 1 ɛ 2 plane, where the bifurcation parameters are defined in Section 5. The analysis for the Hopf case, with small frequencies, is limited to the region close to the ɛ 2 axis as shown. periodic solutions which have maximal symmetry were classified by Roberts, Swift and Wagner [15]. In the Takens Bogdanov case, we recover the same types of periodic solutions, but some of them can now also arise as a secondary bifurcation from steady solutions rather than as a bifurcation directly from the rest state. In some cases, we can also establish the existence of homoclinic or heteroclinic orbits which arise as limits of periodic solutions. The aim of this bifurcation analysis is to provide solutions that are valid far enough from the point of onset to be applicable to experiments. Recall that the analysis for the Hopf bifurcation case [5] is based on assuming a nonzero frequency, and as the frequency tends to zero, the region of validity of the analysis shrinks. Hopf bifurcations close to a Takens Bogdanov point involve eigenvalues with small frequencies, which limits the region of validity of the theory in Fig. 3 to the area close to the negative ɛ 2 axis, while the analysis for the Takens Bogdanov case would apply to the larger shaded region. We identify several physical situations where a Takens Bogdanov point occurs, and evaluate the coefficients of the amplitude equations for each of them. We then look for stable patterns by integrating the amplitude equations in time. In most cases, we find steady solutions on one side of the Takens Bogdanov point and time-periodic solutions on the other side with a transition regime in between. The preferred steady patterns are rolls and hexagons, and the preferred periodic patterns are traveling rolls and wavy rolls (1), and, in one particular case, oscillating triangles [15]. We note that these were also the patterns predicted in the analysis of the Hopf bifurcation [5]. We do not find a transition directly from steady to periodic behavior; instead we observe a variety of quasiperiodic and chaotic solutions in the transition regime. 2. Double-layer convection Fig. 4 illustrates the problem configuration. The upper boundary at z = l is kept at a constant temperature θ0, and the lower boundary at z = 0 is kept at a higher constant temperature θ1 = θ 0 + θ. At the temperature of the top plate, fluid i has coefficient of cubical expansion ˆα i, thermal diffusivity κ i, thermal conductivity k i, viscosity µ i, density ρ i, and kinematic viscosity ν i = µ i /ρ i. S is the dimensional interfacial tension coefficient. The average height of the lower fluid (fluid 1) is l1. The average height of the upper fluid (fluid 2) is l l1 = l 2. Length is made dimensionless with the plate separation l, time with l 2 /κ 1, velocity with κ 1 /l, and pressure with (ρ 1 κ1 2)/l 2. The dimensionless interface position is denoted by z = l 1 + h(x,y,t), velocity by v = (u, v, w), the pressure by p and the temperature by θ. The dimensionless temperature is denoted by θ = (θ θ0 )/ θ (asterisks denote dimensional quantities).

174 Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 Fig. 4. Problem definition. There are six independent dimensionless ratios arising from the stratification in the fluid properties: m = µ 1 /µ 2, r = ρ 1 /ρ 2,γ = κ 1 /κ 2,ζ = k 1 /k 2,β= ˆα 1 / ˆα 2,l 1 = l1 /l = 1 l 2, where l 2 = l2 /l. There are four more independent parameters: a Rayleigh number R = g ˆα 1 θ l 3 /(κ 1 ν 1 ), a Prandtl number P = ν 1 /κ 1, a dimensionless measure of the temperature difference between the plates ˆα 1 θ which should be sufficiently small to be consistent with the Oberbeck Boussinesq approximation, and a surface tension parameter S = S l /(κ 1 µ 1 ). Here, ˆα 1 θ = RP/G, where G = g(l ) 3 /κ1 2. The gravity parameter G appears in the base pressure field, and enters the analysis through the interfacial normal stress condition if there is a density jump across the interface. In each fluid, the governing equations are the heat transport equation, the Navier Stokes equation and incompressibility. The boundary conditions are zero velocity and constant temperature: θ = 1atz = 0, θ= 0atz = 1. The interface is at z = l 1 + h(x,y,t). The conditions to be satisfied at the interface are continuity of velocity, temperature and heat flux, and balance of tractions. 2.1. Base Solution A base solution to the problem is given by h = 0, v = 0, θ = { 1 A1 z for 0 z l 1, A 2 (1 z) for l 1 z 1, (1) A 1 = 1/(l 1 + ζl 2 ), A 2 = ζa 1. Note the corresponding pressure field is found in Joseph and Renardy [16] and enters into the interface continuity conditions in the perturbation equations. 2.2. Perturbation equations We denote by θ the difference between θ and the base solution (1), and by p the difference between p and the base solution. The perturbation stress tensor is denoted T T T = P [ v+( v) T ] p1 in fluid 1, and (P /m)[ v+( v) T ] p1 in fluid 2, where ( v) T denotes the transpose of v. The governing equations are, for 0 z l 1 +h(x,y,t)(layer 1), the heat transport equation θ A 1 w θ = (v ) θ, the momentum equation v P v+ p RP θe z = (v )v, and for l 1 + h(x,y,t) z 1 (layer 2), θ A 2 w θ/γ = (v ) θ, v (r/m)p v + r p (RP /β) θe z = (v )v, together with incompressibility div v = 0. The boundary conditions at the walls z = 0 and z = 1 are v = 0, θ = 0. Let t 1,t 2 denote two unit tangent vectors and let n be a unit normal to the interface. By [[ ]] we denote the jump of a quantity across the interface, i.e. its value in fluid 1 minus its value in fluid 2. The interface conditions at z = l 1 + h are the continuity of velocity, shear stress, the jump in the normal stress is balanced by surface tension and curvature, continuity of temperature, continuity of heat flux, and the kinematic condition. These

Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 175 are, respectively, [[v]] = 0, [[t i T T T n]] = 0, i = 1, 2, (2) { [ ( ) PS 2 2 ] [ ( ) 2 ] } h 1 + h [[n T T T n]] = M 1 h + M 2 h 2 x + 2 y + 2 h 1 + h y 2 x 2 2 h x y x h h y [1 + ( h) 2 ] 3/2, [[ θ]] = h(a 1 A 2 ), [[kn θ]] = 0, ḣ + u h x + v h y = w, M 1 = G ( 1 1 ) ( ) 1 + RPA 2 l 2 r rβ 1, M 2 = RP 2 ( A2 rβ A 1 Terms up to third order are required in the derivation of the amplitude equations. The momentum and heat transport equations contribute quadratic nonlinearities. The interface conditions (2), expanded in Taylor series about the known position z = l 1 and truncated, yield quadratic and cubic nonlinearities. These are lengthy expressions and can be found in ([16] Chapter III) and [2]. We denote by the solution vector (u, v, w, p, θ,h) and write the system of equations, boundary and interface conditions in the schematic form L( ) = N 2 (, ) + N 3 (,, ), (3) where L, N 2,N 3 represent the linear, quadratic and cubic operators, respectively. We perform a three-dimensional bifurcation analysis. ). 3. Examples of Takens Bogdanov points The examples investigated in this section are taken from [4,5]. The work in [5] was motivated by experiments on Fluorinert lying under silicone oil Rhodorsil 47v10. A fictitious fluid, Anderinert, was inadvertently created by a software package which was used by the experimental group to convert units. Anderinert has a thermal diffusivity 3.24 times that of Fluorinert. As it turns out, the onset of instability in the Fluorinert/silicone oil system is always to a steady mode, while the Anderinert system has a window of parameters where a Hopf onset occurs. The coalescence of two real eigenvalues to form a complex conjugate pair as shown in Fig. 2 occurs in both systems, but for the Fluorinert system it only happens after the system has already become unstable, and oscillatory onset does not happen unless the spatial period of solutions is artificially constrained. Indeed, the experiments of [8] show a steady bifurcation occurring first, and then periodic solutions arise from a secondary bifurcation. In both the Fluorinert/silicone oil and Anderinert/silicone oil systems, the onset of instability is close to a Takens Bogdanov point if l 1 is chosen appropriately. Below, we shall pursue the parameters for both of these systems, as well as those used in [4]. The fluid parameters for the Anderinert/silicone oil system are given in row (a) of Table 1. Fig. 5 shows the Rayleigh number and wave number at which onset of instability occurs as functions of l 1. Oscillatory onset occurs in the interval 0.486 l 1 0.504. The endpoints of this interval are not Takens Bogdanov points, because the onset of instability shifts to a different wave number: at l 1 = 0.486, there is an interaction of a real mode with wavenumber 6 and Hopf mode with wavenumber 5.3. At l 1 = 0.504, there is an interaction of a real mode with wavenumber 5.6 and a Hopf mode with wavenumber 5.2. We find Takens Bogdanov points if we fix the wavenumber, say at 5.3, and this is illustrated in Fig. 6. The detailed numerical values for the left and right hand ends of the band of Hopf modes in Fig. 6 are given in Table 2. We note that throughout the interval where Hopf modes exist, the frequencies of the oscillatory modes remain small. In dimensionless terms, the largest frequencies we found are around 5 (for

176 Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 Table 1 (a) The first row gives the fluid parameters for Anderinert/silicone oil. For interfacial tension, a value of 20 in CGS units is used. (b) The second row gives the parameters for Fluorinert/silicone oil. (c) The third row gives parameters for our. description of the example used in Fig. 2 of [17]. In their notation, (α, κ, ρ, λ, a) correspond to our (β,γ,r,ζ,l 2 /l 1 ). Rl 4 1 is their Rayleigh number, their wavenumber k 0 is αl 1 P G S γ β r ζ m (a) 125.3 1.571 10 9 8.3 10 4 1.299 0.926 2.092 0.5385 2.929 (b) 406.3 1.65 10 10 2.7 10 5 0.401 0.926 2.092 0.5385 2.929 (c) 10 5 10 10 10 3 2 0.2 10 1 1 Fig. 5. Onset conditions for the Anderinert/silicone oil system with fluid parameters as in row (a) of Table 1. Fig. 6. Onset conditions for row (a) of Table 1 when the wavenumber is fixed at α = 5.3.

Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 177 Table 2 The values of the lower fluid depth l 1, Rayleigh numbers R and the least stable eigenvalues σ for the Anderinert/silicone oil system are given for points close to the left and right hand ends of the Hopf onset range of Fig. 6, wavenumber 5.3, showing the sensitivity with respect to the parameters. In the cases below, as R increases, complex conjugate eigenvalues turn into reals l 1 R σ 0.48417472 14190.4625 0.0001±0.0006i 0.48417472 14190.4629 0.00008 0.48417472 14190.4629 0.0003 0.50487 13883.1828 0.0035±0.0037i 0.50487 13883.1853 0.00002& 0.00692 Fig. 7. Fluorinert and silicone oil system with l 1 = 0.423,R = 26 498.125. The Takens Bogdanov point is close to α = 4.711. Table 3 Values of the wavenumber, Rayleigh number and least stable eigenvalues are given for the Fluorinert system, second row of Table 1, with l 1 = 0.423 fixed α R σ 4.711 26498.125 0.006±0.0178i 4.71 26502.5 0.003±0.066i instance in Fig. 6 the maximum value is between 3 and 4), in contrast α 2 is around 25 (note that the heat equation involves the combination θ θ). Hence the assumption of small frequency is actually satisfied throughout the range where Hopf modes exist. The Takens Bogdanov analysis carried out in this paper applies near either end of this range; a more comprehensive treatment would need to consider the merger of two Takens Bogdanov points which leads to a more degenerate and complicated bifurcation (see [9,10]). The fluid parameters for the Fluorinert/silicone oil system are given in the second row of Table 1. If the wave number is unconstrained, the onset of instability is always due to real eigenvalues. However, if we fix the spatial period, there are windows of Hopf onsets and Takens Bogdanov points in this system as well. This is illustrated in Fig. 7, which shows the eigenvalues as a function of wavenumber for a fixed Rayleigh number and depth ratio numerical values for the Takens Bogdanov point are given in Table 3. The third row of Table 1 shows data modeling Fig. 2 of [4]. They give the thermal diffusivity ratio (their κ), m/(rβ) (their ν/α), m, the conductivity ratio, their wavenumber k 0 = 2.7, volume ratio a = l 2 /l 1. Our wavenumber α is k 0 /l 1, their a is our l 2 /l 1, their Rayleigh number is Rl1 4. In their problem, the interface is constrained to remain flat and P =. In order to model their problem we choose appropriately large values for our interfacial tension parameter, Prandtl number and G. With our choice of parameters, we examined the second row of their Table 1, where they give l 1 = 0.5, their Rayleigh number 1173, frequency w 0 = 1.29 = Im σl1 2. This translates to R = 18 768, Im σ = 5.16. We find Im σ = 5.26 and neutral R = 18 777. The variation with respect to wavenumbers was examined at R = 18 768 and the maximum growth rate was attained at α = 5.2 rather than 5.4. The small

178 Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 Table 4 The parameters for the third row of Table 1, with α = 5.4 fixed l 1 R σ 0.48303 19560.6 0.03±0.006i 0.4837929 19516.5 0.0087±1.54i 0.5199 17850 0.006±0.135i 0.52 17825.390625 0.04±0.03i 0.52 17825 0.04±0.08i difference between our results and theirs may be due to the fact that they use an approximate set of equations for the limit of infinite Prandtl number (Table 4). 4. Bifurcation analysis at a Hopf bifurcation We give a brief synopsis leading up to the amplitude equation for solutions in the neighborhood of a Hopf onset, with double periodicity on the hexagonal lattice. These results have appeared in [2,5,16], where the final amplitude equations are given in Birkhoff normal form. In this form, several terms including quadratic terms have been transformed away under suitable coordinate transformations and the amplitude equation is then in its simplest form. The Birkhoff normal form of the amplitude equations is the starting point for the analysis of [15], where a comprehensive analysis of bifurcating periodic solutions with maximal symmetry is given. Our analysis for the Takens Bogdanov case will begin with the amplitude equations that result just prior to the transformation to Birkhoff normal form, because it is precisely these coordinate transformations that break down when the frequency tends to zero. In the Takens Bogdanov case, a straightforward calculation of the generalized eigenfunctions is lengthy so we wish to avoid this; we can find them more efficiently by taking a suitable limit of the Hopf eigenfunctions as the problem approaches the Takens Bogdanov case. We return to this issue in Section 5. 4.1. Spatial periodicity on the hexagonal lattice In the x- and y-directions, the solution is assumed to be doubly periodic, e.g., θ(x + n 1 x 1 + n 2 x 2,t) = θ(x,t) for every pair of integers (n 1,n 2 ), where the vectors x 1 and x 2 span a hexagonal lattice of period W : x 1 = W ( 3/2, 1/2, 0), x 2 = W (0, 1, 0). The lattice obtained from this double periodicity is invariant under the symmetries of the hexagon; that is, rotation by multiples of 60, reflection across the vectors a i defined by a 1 = (4π/W 3)(1, 0, 0), a 2 = (4π/W 3)( 1/2, 3/2, 0), a 3 = a 1 a 2, and reflection across axes perpendicular to the a i. The same periodicity condition holds for θ, v, h and p. These variables are expanded in Fourier series; e.g., θ(x,y,z,t) is the sum over k and l of modes θ kl (z, t)e ika 1 x+ila 2 x. We next recall the results for a Hopf bifurcation, show the amplitude equation for this case, and then later build the connection to the Takens Bogdanov case. For the linearized problem (Section 3), the method of separation of variables yields θ kl = e σt θ(z), and similarly for v, p, h. This leads to an eigenvalue problem for σ, in which the results do not depend on the direction of the vector ka 1 + la 2, but on its magnitude α given by ka 1 + la 2 =(4π/ 3W )(k 2 + l 2 kl) 1/2 ; the wavenumber α denotes the critical value determined in Section 3. This equation determines the period W of the lattice. The factor k 2 + l 2 kl can be 0, 1, 3, 4, 7,... The mean flow mode k = l = 0 is not of interest in the linear problem, but will enter into the nonlinear interactions. The smallest nonzero value of k 2 + l 2 kl is 1, for which W is (4π/ 3α). This occurs for six possible pairs (k, l) : (±1, 0), (0, ±1), (1, 1), and ( 1, 1), yielding a sixfold degeneracy of the corresponding eigenvalue. We pursue this case where the lattice size fits exactly into the critical period, and look

Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 179 Fig. 8. The eigenvectors ζ i denote waves traveling to the six vertices of a hexagon constructed with vectors a 1,a 2,a 3, a 1, a 2, a 3. The corresponding wave amplitudes are z i. at the nonlinear interaction generated by the six Fourier modes. (If k 2 + l 2 kl is chosen larger than one, then we would seek solutions on a larger lattice with period a multiple of the critical one.) At criticality, there is a complex conjugate pair of eigenvalues σ =±iω which has six eigenfunctions denoted by ζ k, corresponding to the six values of (k, l). The eigenfunctions computed in Section 3 correspond to the case of l = 0, k= 1 and are denoted by ζ 1 and of the form ζ(z)exp(ia 1 x), where a 1 = (α c, 0, 0). The vectors ±a 1, ±a 2 and ±a 3 emanate from the center of a hexagon and terminate at its six vertices. The critical eigenfunctions are waves propagating in the directions of the vertices. Fig. 8 illustrates the eigenvectors. Let the parameter λ denote a set of bifurcation parameters, e.g., the difference between the Rayleigh number and its critical value R R C, or one of the fluid property ratios. Close to criticality, an initial disturbance proportional to ζ k (λ) evolves as exp[ µ(λ)t]. At criticality, λ = 0, µ(0) = iω. We denote the complex time-dependent amplitude function of the wave propagating in the direction of a i by z i. In order to obtain an amplitude evolution equation for the weakly nonlinear analysis, the governing equations are reduced to a system of ordinary differential equations in the 12-dimensional space R 12 by invoking the center manifold theorem which says roughly that in the neighborhood of criticality, the dynamics is governed by interactions among the six critical modes. This allows us to write the solution in the form = 1 + 2 + higher order terms, (4) 6 6 6 1 = z i ζ i + z i ζ i, 2 = 2Re z i z j ψ ij + z i z j χ ij, i=1 i=1 i,j=1 where the interaction terms ψ ij,χ ij are found by certain projections in the center manifold reduction scheme. Details of the derivation are contained in [2,16]. The following amplitude equation is obtained at the bifurcation point: dz i + F i (z 1,z 2,z 3,z 4,z 5,z 6,λ)= 0, i = 1,...,6, (5) dt where F i,i = 2,...,6 can be retrieved from F 1 by permutation of the arguments: F 2 (z 1,z 2,z 3,z 4,z 5,z 6 ) = F 1 (z 2,z 3,z 1,z 5,z 6,z 4 ), F 3 (z 1,z 2,z 3,z 4,z 5,z 6 ) = F 1 (z 3,z 1,z 2,z 6,z 4,z 5 ), F 4 (z 1,z 2,z 3,z 4,z 5,z 6 ) = F 1 (z 4,z 5,z 6,z 1,z 2,z 3 ), F 5 (z 1,z 2,z 3,z 4,z 5,z 6 ) = F 1 (z 5,z 6,z 4,z 2,z 3,z 1 ), F 6 (z 1,z 2,z 3,z 4,z 5,z 6 ) = F 1 (z 6,z 4,z 5,z 3,z 1,z 2 ), (6)

180 Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 and F 1 (z 1,z 2,z 3,z 4,z 5,z 6,λ)= µ(λ)z 1 + β 1 (λ)z 5 z 6 + β 2 (λ) z 2 z 3 + β 3 (λ)(z 5 z 3 + z 6 z 2 ) + γ 1 (λ) z 1 2 z 1 +γ 2 (λ) z 4 2 z 1 + γ 3 (λ)z 2 1 z 4 + γ 4 (λ) z 2 4 z 1 + γ 5 (λ) z 1 2 z 4 + γ 6 (λ) z 4 2 z 4 +γ 7 (λ)( z 2 2 + z 3 2 )z 1 + γ 8 (λ)( z 2 2 + z 3 2 ) z 4 + γ 9 (λ)( z 5 2 + z 6 2 )z 1 +γ 10 (λ)( z 5 2 + z 6 2 ) z 4 + γ 11 (λ)(z 2 z 5 + z 3 z 6 )z 1 + γ 12 (λ)(z 2 z 5 + z 3 z 6 ) z 4 +γ 13 (λ)( z 2 z 5 + z 3 z 6 )z 1 + γ 14 (λ)( z 2 z 5 + z 3 z 6 ) z 4. (7) Terms of higher than third degree have been ignored. We remark that, in order to remove certain degeneracies, it is necessary to include some fifth-order terms in the analysis of the Hopf bifurcation [15]. This is not the case for the Takens Bogdanov analysis, because the degeneracy is also removed by quadratic terms. These quadratic terms transform away in the Birkhoff normal form for the Hopf bifurcation, but this transformation breaks down in the Takens Bogdanov limit. The coefficients β i and γ i are the Landau constants defined in [2] and in ([16], Chapter III.7). The numerical code for calculating the Landau coefficients is developed from that used in [2,5]. We comment in passing that, in the process, an error has been detected in the calculation of χ 15 and χ 24. This affects the value of the Landau coefficient α 4 in [2,5]. It turns out that the qualitative nature of results in [2] is unaffected. The problem studied in [2] is the two-layer Bénard problem in the case where oscillatory onset arises due to the coupling between the Bénard instability and the interface motion; it is found in this case that the 11 solutions found in [15] are unstable for the three data considered in [2]. In the present problem, where oscillatory onset arises due to the coupling of the motions in both layers, with negligible interface motion, the results are different, and stable periodic branches are found to exist. The revised results for [5] are contained in [18]. The results indicate that the periodic solutions most likely to be stable are the traveling rolls and the wavy rolls (1). 5. Bifurcation analysis at a Takens Bogdanov point In the linearized stability analysis, we focus on an onset condition with a double zero eigenvalue. Associated with this, there are six eigenfunctions denoted by ζ k, and six generalized eigenfunctions denoted by ξ k. We consider this as a limiting case of the Hopf bifurcation analysis. As an example of the limits, we consider the case of the 2 2 matrices D 1 = ( ) 0 1, D 2 = 0 0 ( ) 0 1. ɛ 0 The second matrix is a perturbation of the first; D 1 has a double zero eigenvalue, and D 2 (the Hopf case) has complex eigenvalues µ = i ɛ and µ = i ɛ. The eigenvectors for the Hopf case are ( ) ( ) 1 1 i and ɛ i. ɛ In analogy with the present problem, we refer to these as ζ 1 and ζ 4, respectively. As ɛ 0,ζ 1 and ζ 4 become the same and the set of eigenvectors is no longer linearly independent. D 1 has eigenvector ( 1 0 ). The generalized eigenvector ξ satisfies D 1 ξ = the eigenvector, so ξ = ( 0 1 ). The eigenvector is (ζ 1 + ζ 4 )/2 and the generalized eigenvector is (ζ 1 ζ 4 )/(µ µ). Note that as ɛ 0, the latter approaches a finite limit due to the expression

Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 181 µ µ in the denominator. Thus, (ζ 1 + ζ 4 )/2 and (ζ 1 ζ 4 )/(µ µ) provide a linearly independent basis for the eigenspace in the limit as ɛ 0. We return to the Hopf bifurcation on the hexagonal lattice, and consider the component of the solution 1 proportional to exp(iαx). We express this in terms of the eigenvector and generalized eigenvector of the Takens Bogdanov problem: z 1 ζ 1 + z 4 ζ 4 = v 1 ζ 1 + ζ 4 2 w 1 ζ 1 ζ 4 µ µ. (8) The amplitude functions for the Takens Bogdanov case are v 1 and w 1. From this, we see that z 1 = v 1 /2 w 1 /(µ µ), z 4 = v 1 /2 + w 1 /(µ µ), z 2 = v 2 /2 w 2 /(µ µ), z 5 = v 2 /2 + w 2 /(µ µ), z 3 = v 3 /2 w 3 /(µ µ), and z 6 = v 3 /2 + w 3 /(µ µ). The amplitude equations (7) have the form ż 1 + µz 1 = nonlinear terms, and z 4 + µ z 4 = nonlinear terms (this is obtained from (7) with z 1 z 4,z 2 z 5,z 3 z 6 ). Our aim is to obtain amplitude equations for the v 1 and w 1. Therefore, we add the ż 1 -and z 4 -equations, and use z 1 + z 4 = v 1 and µz 1 + µ z 4 = [(µ+ µ)/2]v 1 w 1, to obtain v 1 +[(µ+ µ)/2]v 1 w 1 = nonlinear terms. We subtract the equations, multiply through by (µ µ)/2, and obtain ẇ 1 ((µ µ) 2 /4)v 1 + ([µ + µ]/2)w 1 = nonlinear terms. We define bifurcation parameters ɛ 1 = (µ + µ)/2 and ɛ 2 = (µ µ) 2 /4. Our previously defined notation λ represents the set (ɛ 1,ɛ 2 ). On the Hopf side of the Takens Bogdanov point, µ(0) = iω so ɛ 1 = 0, ɛ 2 = ω 2 (ω 0). The base solution is unstable for ɛ 1 < 0 (since ζ k exp( µ(λ)t)). This leads to amplitude equations as follows (up to cubic order) 0 =ẇ 1 + ɛ 1 w 1 ɛ 2 v 1 + b 1 w 2 w 3 + b 2 v 2 v 3 + b 3 ( w 2 v 3 + w 3 v 2 ) + c 1 w 1 2 w 1 + c 2 w 1 2 v 1 +c 3 v 1 2 w 1 + c 4 v 1 2 v 1 + c 5 v 2 1 w 1 + c 6 w 2 1 v 1 + c 7 ( v 2 2 + v 3 2 )v 1 + c 8 ( v 2 2 + v 3 2 )w 1 +c 9 (v 2 w 2 + v 3 w 3 )v 1 + c 10 ( v 2 w 2 + v 3 w 3 )v 1 +..., (9) 0 = v 1 w 1 + ɛ 1 v 1 + b 1 w 2 w 3 + b 2 v 2 v 3 + b 3 ( w 2 v 3 + w 3 v 2 ) + c 1 w 1 2 w 1 + c 2 w 1 2 v 1 + c 3 v 1 2 w 1 + c 4 v 1 2 v 1 + c 5 v 2 1 w 1 + c 6 w 2 1 v 1 + c 7 ( v 2 2 + v 3 2 )v 1 + c 8 ( v 2 2 + v 3 2 )w 1 + c 9 (v 2 w 2 + v 3 w 3 )v 1 + c 10 ( v 2 w 2 + v 3 w 3 )v 1 +..., (10) where the coefficients b i,c i, b i, c i are real. The remaining equations are found by cyclic permutation of the indices. We can transform away the ɛ 1 -term in Eq. (10) by the substitution ŵ 1 = w 1 ɛ 1 v 1. This yields the new equations 0 =ẇ 1 + 2ɛ 1 w 1 ɛ 2 v 1 + b 1 w 2 w 3 + b 2 v 2 v 3 + b 3 ( w 2 v 3 + w 3 v 2 ) + c 1 w 1 2 w 1 + c 2 w 1 2 v 1 +c 3 v 1 2 w 1 + c 4 v 1 2 v 1 + c 5 v 2 1 w 1 + c 6 w 2 1 v 1 + c 7 ( v 2 2 + v 3 2 )v 1 + c 8 ( v 2 2 + v 3 2 )w 1 +c 9 (v 2 w 2 + v 3 w 3 )v 1 + c 10 ( v 2 w 2 + v 3 w 3 )v 1 +..., (11) 0 = v 1 w 1 + b 1 w 2 w 3 + b 2 v 2 v 3 + b 3 ( w 2 v 3 + w 3 v 2 ) + c 1 w 1 2 w 1 + c 2 w 1 2 v 1 + c 3 v 1 2 w 1 + c 4 v 1 2 v 1 + c 5 v 2 1 w 1 + c 6 w 2 1 v 1 + c 7 ( v 2 2 + v 3 2 )v 1 + c 8 ( v 2 2 + v 3 2 )w 1 + c 9 (v 2 w 2 + v 3 w 3 )v 1 + c 10 ( v 2 w 2 + v 3 w 3 )v 1 +... (12) Here we have suppressed the hats, as well as terms which represent O(ɛ) perturbations to the coefficients. The following analysis will proceed on the basis of Eqs. (11) and (12). The small parameters ɛ 1 and ɛ 2 determine the behavior of the linearized system. We note that the eigenvalues of the linearization are ɛ 1 ± ɛ1 2 + ɛ 2. (13)

182 Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 Thus, we have real eigenvalues if ɛ 2 > ɛ 2 1 and complex eigenvalues if ɛ 2 < ɛ 2 1. Moreover, ɛ 1 is the average of the two eigenvalues, and in the case of complex eigenvalues, the onset of instability (Hopf bifurcation) occurs when ɛ 1 = 0. 5.1. Rescaling of the equations The aim of introducing new scales is to balance the linear terms with some of the nonlinear terms, while keeping the symmetries in the problem as they are. Thus, the v i are scaled in the same manner, and the w i also. According to Eq. (12), we ought to scale v i in the same manner as w i, and Eq. (11) suggests scaling ẇ i in the same way as ɛ 2 v i. We can balance linear and quadratic terms in Eq. (11) if we scale ẇ i in the same fashion as vi 2. For the case where the coefficients of the quadratic terms are small, and also to study the pattern of rolls for which quadratic interaction terms vanish identically, it is of interest to consider a second scaling where ẇ i is scaled in the same way as vi 3. These considerations lead to the following two cases. 5.1.1. Case A Here the coefficients of quadratic terms are assumed to be of order one, and cubic terms enter only as a lower order perturbation. d/dt = ɛd/dτ,v i = ɛṽ i,ɛ 2 = ±ɛ, w i = ɛ 3/2 w i,ɛ 1 = γɛ. To simplify notation, we have written v i,w i for ṽ i, w i in the equations that follow, and we now use a dot to denote d/dτ. This leads to v 1 w 1 + ɛ b 2 v 2 v 3 = 0 and ẇ 1 v 1 + b 2 v 2 v 3 + ɛ(2γw 1 + b 3 ( w 2 v 3 + w 3 v 2 )) = 0. These combine into one equation: v 1 v 1 + b 2 v 2 v 3 + ɛ[2γ v 1 + ( v 2 v 3 + v 3 v 2 )( b 2 + b 3 )] = 0. 5.1.2. Case B The coefficients of the quadratic terms are small (this scaling is also appropriate for rolls for which the quadratic terms vanish): d/dt = ɛd/dτ,v i = ɛṽ i,w i = ɛ 2 w i,b i = ɛ ˆb i, b i = ɛ ˆ b i,ɛ 2 =±ɛ 2,ɛ 1 = γɛ 2. v 1 w 1 + ɛ(ˆ b 2 v 2 v 3 + c 4 v 1 2 v 1 + c 7 ( v 2 + v 3 2 )v 1 ) = 0, ẇ 1 v 1 + ˆb 2 v 2 v 3 + c 4 v 1 2 v 1 + c 7 ( v 2 2 + v 3 2 )v 1 + ɛ[2γw 1 + ˆb 3 ( w 2 v 3 + w 3 v 2 ) + c 3 v 1 2 w 1 + c 5 v1 2 w 1 + c 8 ( v 2 2 + v 3 2 )w 1 + c 9 ( w 2 v 2 + w 3 v 3 )v 1 + c 10 (w 2 v 2 + w 3 v 3 )v 1 )] = 0. (14) Elimination yields: v 1 v 1 + ˆb 2 v 2 v 3 + c 4 v 1 2 v 1 + c 7 ( v 2 2 + v 3 2 )v 1 + ɛ[2γ v 1 + ( ˆb 3 + ˆ b 2 )( v 2 v 3 + v 3 v 2 ) + (c 3 + 2 c 4 ) v 1 2 v 1 + (c 5 + c 4 )v1 2 v 1 + (c 8 + c 7 )( v 2 2 + v 3 2 ) v 1 + (c 9 + c 7 )( v 2 v 2 + v 3 v 3 )v 1 + (c 10 + c 7 )( v 2 v 2 + v 3 v 3 )v 1 ] = 0. (15) In the following, we discuss various types of solutions for these equations. We shall focus on Case B, since first, the equations for Case A are a subset of those for Case B, and second, the equations for Case A are too severely truncated to capture important aspects of the dynamics; for instance, they do not allow for rolls solutions or for stable steady hexagons. In the following discussion, we shall drop the hat on the b i and b i in Eq. (15). We note that for ɛ = 0, the Eq. (15) is a Hamiltonian system with Hamiltonian H = 1 3 v i 2 1 3 v i 2 + b 2 Re( v 1 v 2 v 3 ) + 1 3 2 2 4 c 4 v i 4 + 1 2 c 7( v 1 2 v 2 2 + v 2 2 v 3 2 + v 3 2 v 1 2 ). (16) i=1 i=1 i=1

Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 183 (Denote v i = x i + iy i, then H = (1/2) k p2 k + V(q 1,...,q 6 ), where the q k denote x 1,y 1,...,x 3,y 3, q k = p k, the p k denote ẋ 1,...,ẏ 3, ṗ k = V / q k.) It would be possible to subject Eq. (15) to further transformations to achieve a normal form with fewer coefficients involved. We shall not pursue this, since our discussion below focusses on the Hamiltonian system obtained for ɛ = 0, which cannot be simplified further. 5.2. Time-periodic patterns Our analysis will be concerned mostly with periodic solutions. In contrast to the case of Hopf bifurcation, however, we are not concerned just with the bifurcation of periodic solutions from the rest state. A new possibility in the Takens Bogdanov case is the secondary bifurcation of periodic solutions from steady states, which we shall investigate in some detail. Other new features include quasiperiodic states and homoclinic or heteroclinic orbits which can arise as limiting cases of periodic orbits. The patterns found in [15] for the Hopf bifurcation will play an important role in our analysis as well as in the numerical results below. The following is a list of possible steady solutions and of the periodic solutions of [15], together with the symmetries which they satisfy. We note that the translation symmetries of the original problem manifest themselves in Eq. (15) as the symmetry v 1 v 1 exp(2iα), v 2 v 2 exp( iα+iβ),v 3 v 3 exp( iα iβ). For the list below, we have placed centers or axes of symmetry conveniently to get the simplest possible form (e.g. hexagons as listed have their center of symmetry at the origin); clearly other solutions (e.g. a two-parameter family of hexagons) can be generated by spatial translations. The possible steady patterns are as follows: 1. Rolls: v 2 = v 3 = 0, v 1 real. 2. Hexagons: v 1 = v 2 = v 3 real. 3. Rectangles (patchwork quilt): v 1 real, v 2 = v 3 real. The latter class of solutions is the one referred to as type IV in [19]. The periodic solutions of Roberts, Swift and Wagner [15] are as follows: 1. Standing rolls: v 2 = v 3 = 0,v 1 real, v 1 (t + T/2) = v 1 (t). 2. Standing hexagons: v 1 = v 2 = v 3 real. 3. Standing patchwork quilt: v 1 real, v 2 = v 3 real, v 2 (t + T/2) = v 2 (t), v 1 (t + T/2) = v 1 (t). 4. Standing regular triangles: v 1 = v 2 = v 3,v 1 (t + T/2) = v 1 (t). 5. Traveling rolls: v 2 = v 3 = 0,v 1 = V exp(iωt). 6. Traveling patchwork quilt (1): v 1 = V 1 exp(2iωt), v 2 = v 3 = V 2 exp( iωt). 7. Traveling patchwork quilt (2): v 1 = V 1 real, v 2 = v 3 = V 2 exp(iωt). 8. Oscillating triangles: v 1 = v 2 = v 3,v 1 (t + T/3) = exp(2πi/3)v 1 (t). 9. Twisted patchwork quilt: v 1 real, v 2 (t) = v 1 (t + T/3), v 3 (t) = v 1 (t + 2T/3). 10. Wavy rolls (2): v 1 (t + T/2) = v 1 (t), v 2 (t) = v 1 (t + T/3), v 3 (t) = v 1 (t + 2T/3). 11. Wavy rolls (1): v 1,v 2,v 3 real, v 2 (t) = v 1 (t + T/4), v 1 (t) = v 2 (t + T/4), v 3 (t + T/4) = v 3 (t). To look for periodic solutions of Eq. (15), we proceed in two steps. We first set ɛ = 0 and look for periodic orbits of the Hamiltonian system. These will exist in one-parameter families (once we fix centers or axes of symmetry to eliminate the translation symmetries). When dissipation is added in (ɛ > 0), only isolated solutions in each oneparameter family persist. In nondegenerate cases, persistence is determined by an integral condition requiring the average dissipation to be zero. For each periodic orbit of the Hamiltonian system, this yields a condition determining γ as a function of ɛ if the other coefficients are given. Specifically, the condition for persistence of periodic orbits is obtained as follows (this discussion is analogous to Lemma 2.1, on p. 445 of [20]). With q k defined as in the remark following Eq. (16), we can put Eq. (15) in the

184 Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 form q k = V q k + ɛf k (q). (17) For ɛ = 0, let q 0 (t) be a periodic solution with period T. If we then look for periodic solutions for small ɛ by a regular perturbation expansion, then the solvability condition T q k 0 (t)f k(q 0 (t)) dt = 0 (18) 0 k arises at the leading order. Since F k (q) = 2γ q k +..., (19) with the dot indicating terms which do not depend on γ, we can use Eq. (18) to solve for γ. 5.3. Secondary bifurcation of periodic orbits from steady solutions We shall now look for branches of periodic solutions which bifurcate from steady solutions. We focus our analysis on the Hamiltonian system which is obtained by setting ɛ = 0 in Eq. (15): v 1 v 1 + b 2 v 2 v 3 + c 4 v 1 2 v 1 + c 7 ( v 2 2 + v 3 2 )v 1 = 0, v 2 v 2 + b 2 v 1 v 3 + c 4 v 2 2 v 2 + c 7 ( v 1 2 + v 3 2 )v 2 = 0, v 3 v 3 + b 2 v 2 v 1 + c 4 v 3 2 v 3 + c 7 ( v 2 2 + v 1 2 )v 3 = 0. (20) For steady solutions, the system (20) is exactly the same which is found in the study of steady onset for the Bénard problem (without up down symmetry, see [19]). With the equations truncated at the cubic level as in Eq. (20), there are three types of steady solutions: rolls, hexagons, and solutions with rectangular symmetry, which we shall refer to as steady patchwork quilt (they are called type IV solutions in [19]). The latter solutions do not bifurcate directly from the trivial state unless b 2 = 0, instead they form a secondary branch which connects rolls and hexagons. 5.3.1. Bifurcation from steady rolls Steady rolls are given by v 1 = V,v 2 = v 3 = 0, where V is real and V + c 4 V 3 = 0. (21) The linearization of Eq. (20) at a steady roll solution leads to the problem (with y i denoting the linearized perturbation to v i ) ÿ 1 + c 4 V 2 (y 1 +ȳ 1 ) = 0, ÿ 2 + b 2 V ȳ 3 + (c 7 c 4 )V 2 y 2 = 0, ÿ 3 + b 2 V ȳ 2 + (c 7 c 4 )V 2 y 3 = 0. (22) We are interested in periodic solutions with a time dependence proportional to exp(iωt). The eigenvalue problem for ω leads to the following eigenspaces: 1. y 1 = ȳ 1,y 2 = y 3 = 0 and ω = 0. This zero eigenvalue results from the translation symmetry of the problem, which allows for a translation of the steady rolls pattern. Associated with this translation symmetry, we have

Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 185 traveling rolls bifurcating from the one-parameter family of steady rolls. We shall find traveling wave solutions explicitly in the next subsection. The abstract situation for bifurcation of traveling waves is as follows: we have a system of equations of the form ü = F (u), u R 2n, (23) where F is invariant under an O(2)-symmetry, i.e. there is a group of rotation exp(lτ) and a reflection R such that F(exp(Lτ)u) = exp(lτ)f (u), F (Ru) = RF (u), RL = LR. (24) Let u 0 be a steady solution invariant under reflection, but not rotation, i.e. Ru 0 = u 0, Lu 0 0. The space R 2n can be decomposed into an even subspace (Ru = u) and an odd subspace (Ru = u), both these subspaces are invariant under the Jacobian DF(u 0 ).WehaveDF(u 0 )Lu 0 = 0, and Lu 0 is in the odd subspace; generically, there are no null vectors of DF(u 0 ) in the even subspace. Now look for solutions of Eq. (23) which are traveling waves, i.e. they have the form exp(αlt)v, where α is a constant and v does not depend on t. We impose the additional condition that v is even: Rv = v. Then Eq. (23) yields α 2 L 2 v = F(v), (25) and the implicit function theorem yields a unique solution v = v(α 2 ) in a neighborhood of α = 0,v = u 0. 2. y 1 = ȳ 1,y 2 = y 3 = 0 and ω 2 = 2c 4 V 2.Ifc 4 > 0, then there exists a family of periodic orbits of the Hamiltonian system, which are standing rolls, i.e. y 1 is real and y 2 = y 3 = 0. These standing rolls are not the standing rolls of [15], since they oscillate about a nonzero mean. In the analysis of [21], these are the standing rolls of type SW 3. 3. y 1 = 0,y 2 = ȳ 3 and ω 2 = b 2 V + (c 7 c 4 )V 2, This eigenspace is two-dimensional. We have an O(2) symmetry, with translation given by y 2 y 2 exp(iα), y 3 y 3 exp( iα) and reflection given by exchange of y 2 and y 3.Ifω 2 > 0, the nonlinear problem will therefore allow for standing and traveling wave solutions (see [22], Chapter XVII). This leads to a family of periodic orbits which are standing patchwork quilts: v 2 = v 3 real, and a family of traveling patchwork quilts (2): v 2 (t + τ) = exp(iωτ)v 2 (t). 4. y 1 = 0,y 2 = ȳ 3 real and ω 2 = b 2 V + (c 7 c 4 )V 2. This case is equivalent to the situation obtained when y 2 = ȳ 3 and the sign of V is reversed (note that Eq. (15) is invariant under simultaneous sign change of v 1 and v 3, and that this transformation can be interpreted as a translation). Hence the periodic orbits which exist for ω 2 > 0 are again a family of standing patchwork quilts and a family of traveling patchwork quilts (2). 5.3.2. Secondary bifurcation from steady hexagons Steady hexagons are given by v 1 = v 2 = v 3 = V, where V is real and V + b 2 V 2 + (c 4 + 2c 7 )V 3 = 0. (26) The linearization of Eq. (20) at a steady hexagon leads to the system ÿ 1 + b 2 V(ȳ 2 +ȳ 3 y 1 ) + c 4 V 2 (y 1 +ȳ 1 ) + c 7 V 2 (y 2 +ȳ 2 + y 3 +ȳ 3 ) = 0, ÿ 2 + b 2 V(ȳ 1 +ȳ 3 y 2 ) + c 4 V 2 (y 2 +ȳ 2 ) + c 7 V 2 (y 1 +ȳ 1 + y 3 +ȳ 3 ) = 0, ÿ 3 + b 2 V(ȳ 2 +ȳ 1 y 3 ) + c 4 V 2 (y 3 +ȳ 3 ) + c 7 V 2 (y 2 +ȳ 2 + y 1 +ȳ 1 ) = 0. (27) We have the following eigenspaces for the linearized problem and associated bifurcating branches: 1. y 1 = y 2 = y 3 real and ω 2 = b 2 V + (2c 4 + 4c 7 )V 2.Ifω 2 > 0, there is a branch of standing hexagons.

186 Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 2. y 1 = y 2 = y 3 imaginary and ω 2 = 3b 2 V.Ifω 2 > 0, this leads to a branch of standing regular triangles bifurcating from the steady hexagons. 3. y 1,y 2,y 3 imaginary and y 1 +y 2 +y 3 = 0. This leads to a double zero eigenvalue associated with the translation invariance of the problem. Associated with this double zero eigenvalue, we have bifurcating branches of both types of traveling patchwork quilt, see the next subsection. 4. y 1,y 2,y 3 real, y 1 + y 2 + y 3 = 0, and ω 2 = 2b 2 V + 2c 4 V 2 2c 7 V 2.Ifω 2 > 0, we can use the equivariant Hopf bifurcation theorem for systems with D 3 -symmetry (see Theorem 4.1, p. 390 in [22]) to obtain the following types of bifurcating solutions: (a) Solutions for which two of the amplitudes, say v 2 and v 3 are equal. Such solutions have rectangular symmetry. They are a type of standing patchwork quilt, but not the standing patchwork quilts of [15], since they oscillate about a steady hexagon, so all amplitudes have nonzero mean. (b) Solutions for which v 2 (t) = v 1 (t +T/3), v 3 (t) = v 1 (t +2T/3). In the linearized problem, these solutions are given by y 1 = cos(ωt), y 2 = cos(ωt + 2π/3), y 3 = cos(ωt + 4π/3). These solutions are twisted patchwork quilts. (c) Solutions for which v 2 (t + T/2) = v 3 (t), v 3 (t + T/2) = v 2 (t) and v 1 (t + T/2) = v 1 (t). In the linearized problem, such a solution is given by y 2 = y 3,y 1 = 0. The spatial patterns on this solution branch have point symmetry across the origin (since v 1,v 2 and v 3 are real). Moreover, shift by half a temporal period is equivalent to reflection across the x-axis (i.e., exchange of v 2 and v 3 ). 5.3.3. Secondary bifurcation from steady patchwork quilts Steady patchwork quilts are given by v 1 = U, v 2 = v 3 = V, where U and V are real and U + b 2 V 2 + c 4 U 3 + 2c 7 V 2 U = 0, V + b 2 UV + c 4 V 3 + c 7 (U 2 + V 2 )V = 0. (28) If we exclude the case of rolls (V = 0), we can divide the second equation by V and obtain 1 + b 2 U + c 4 V 2 + c 7 (U 2 + V 2 ) = 0. (29) Next, we subtract the two equations in (28) from each other and divide by U V (the case U = V leads to hexagons). The result is 1 b 2 V + c 4 (U 2 + UV + V 2 ) + c 7 (V 2 UV) = 0. (30) Next, we subtract Eq. (30) from Eq. (29) and divide by U + V (the case U + V = 0 also leads to hexagons). The result is U = b 2. c 4 c 7 From Eq. (29), we then find (31) (c 4 + c 7 )V 2 =±1 c 4b2 2 (c 4 c 7 ) 2. (32) The equations for linearized perturbations are ÿ 1 y 1 + b 2 V(ȳ 2 +ȳ 3 ) + c 4 U 2 (2y 1 +ȳ 1 ) + 2c 7 V 2 y 1 + c 7 UV(y 2 +ȳ 2 + y 3 +ȳ 3 ) = 0, ÿ 2 y 2 + b 2 (Uȳ 3 + V ȳ 1 ) + c 4 V 2 (2y 2 +ȳ 2 ) + c 7 (U 2 + V 2 )y 2 + c 7 UV(y 1 +ȳ 1 ) + c 7 V 2 (y 3 +ȳ 3 ) = 0, ÿ 3 y 3 + b 2 (Uȳ 2 + V ȳ 1 ) + c 4 V 2 (2y 3 +ȳ 3 ) + c 7 (U 2 + V 2 )y 3 + c 7 UV(y 1 +ȳ 1 ) + c 7 V 2 (y 2 +ȳ 2 ) = 0. (33)

Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 187 We can identify the following eigenspaces 1. y 1,y 2,y 3 imaginary, y 2 = y 3,y 1 = 2Uy 2 /V, ω 2 = 0. This zero eigenvalue is associated with the translation symmetry in the x-direction. Associated with this symmetry, we have a bifurcation of traveling patchwork quilts (1). 2. y 2,y 3 imaginary, y 2 = y 3,y 1 = 0,,ω 2 = 0. This zero eigenvalue is associated with the translation symmetry in the y-direction. Associated with this translation symmetry, there is a bifurcating branch of traveling patchwork quilts (2). 3. y 1,y 2,y 3 imaginary, y 2 = y 3,y 1 = Vy 2 /U, ω 2 = b 2 (2U + V 2 /U). A branch of periodic solutions exists if ω 2 > 0. These periodic solutions retain symmetry across the x-axis, i.e. v 2 = v 3, but they become reflected across the y-axis every half period: v 1 (t + T/2) = v 1 (t), v 2 (t + T/2) = v 3 (t), v 3 (t + T/2) = v 2 (t). 4. y 1,y 2,y 3 real, y 2 = y 3,y 1 = 0, ω 2 = 1 b 2 U + 3c 4 V 2 + c 7 (U 2 V 2 ). This leads to branches of periodic solutions which preserve point symmetry across the origin, but become reflected across the axes every half period: v 1 (t + T/2) = v 1 (t), v 2 (t + T/2) = v 3 (t), v 3 (t + T/2) = v 2 (t). 5. y 1,y 2,y 3 real, y 2 = y 3. Such solutions preserve the rectangular symmetry. The linearized eigenvalue problem is two-dimensional: ω 2 y 1 y 1 + 2b 2 Vy 2 + 3c 4 U 2 y 1 + 2c 7 V 2 y 1 + 4c 7 UVy 2 = 0, ω 2 y 2 y 2 + b 2 (Uy 2 + Vy 1 ) + 3c 4 V 2 y 2 + c 7 (U 2 y 2 + 2UVy 1 + 3V 2 y 2 ) = 0. (34) In general, there are two different eigenvalues for ω 2, and branches of periodic solutions exist if one or both of these eigenvalues are positive. The patterns resulting from this are standing patchwork quilts with nonzero mean, analogous to those which we found bifurcating from steady hexagons. 5.4. Traveling waves We can identify two types of traveling waves (other than rolls), corresponding to the two types of traveling patchwork quilts in [15]. For the first kind of traveling wave, v 1 is real and time-independent, and v 2 = v 3 is complex and has a time dependence proportional to e iωt : v 1 = V 1, v 2 = V 2 e iωt, v 3 = V 2 e iωt. This type of traveling wave corresponds to the traveling patchwork quilt (2) in [15] and to the traveling wave of type C2 in [10]. For the present case, we obtain the algebraic system V 1 + b 2 V 2 2 + c 4 V 3 1 + 2c 7 V 2 2 V 1 = 0, ω 2 V 2 V 2 + b 2 V 1 V 2 + c 4 V 2 2 V 2 + c 7 ( V 2 2 + V 2 1 )V 2 = 0. (35) We divide the second equation by V 2, and solve for V 2 2 : V 2 2 = ω2 ± 1 b 2 V 1 c 7 V1 2 > 0. (36) c 4 + c 7 We can now insert this result in the first equation of (35), leading to a cubic equation for V 1. We note that this type of traveling wave becomes a steady roll when V 2 = 0. See also the discussion of secondary bifurcations from steady rolls above. The second type of traveling wave corresponds to the traveling patchwork quilt (1) of [15] and the traveling waves of type D of [10]. For these solutions, we have v 1 = V 1 e iωt, v 2 = v 3 = V 2 e iωt/2. (37)

188 Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 We obtain the algebraic system ω 2 V 1 V 1 + b 2 V 2 2 + c 4 V 1 2 V 1 + 2c 7 V 2 2 V 1 = 0, ω2 4 V 2 V 2 + b 2 V 1 V 2 + c 4 V 2 2 V 2 + c 7 ( V 2 2 + V 1 2 )V 2 = 0. (38) We multiply the first equation by V 1, and the second by V 2, resulting in ω 2 V 1 2 V 1 2 + b 2 V 2 2 V 1 + c 4 V 1 4 + 2c 7 V 2 2 V 1 2 = 0, ω2 4 V 2 2 V 2 2 + b 2 V 1 V 2 2 + c 4 V 2 4 + c 7 ( V 2 2 + V 1 2 ) V 2 2 = 0. (39) If b 2 0, we conclude that V2 2V 1 must be real. We set r 1 =± V 1,ρ 2 = V 2 2, with the sign chosen such that V 1 V 2 2 = r 1ρ 2. After dividing the first equation of (39) by r 1 and the second by ρ 2,wefind ω 2 r 1 r 1 + b 2 ρ 2 + c 4 r 3 1 + 2c 7ρ 2 r 1 = 0, ω2 4 1 + b 2r 1 + c 4 ρ 2 + c 7 (ρ 2 + r1 2 ) = 0. (40) We solve the second equation for ρ 2 : ρ 2 = (ω2 /4) ± 1 b 2 r 1 c 7 r1 2 > 0. (41) c 4 + c 7 By inserting Eq. (41) into the first equation of (40), we obtain a third-degree equation in r 1. This type of traveling wave becomes a traveling roll when ρ 2 = 0. We note that if we set ω = 0 for either type of traveling patchwork quilt, we recover the steady patchwork quilts. 5.5. Heteroclinic orbits We can expect a multitude of solutions with more complicated time dependence than periodicity. For instance, the Hamiltonian system for ɛ = 0 will have many invariant tori, and if we include the dissipation, we can expect some of these tori to persist, leading to quasiperiodic solutions. Dangelmayr and Knobloch [21] identified such a family of quasiperiodic solutions for the case of rolls. In the numerical results of the next section, we shall indeed see some quasiperiodic regimes, as well as chaotic ones. For the simple cases of rolls and hexagons, the Hamiltonian system can be integrated, and it is easy to see that homoclinic or heteroclinic orbits exist as limiting cases of periodic solutions. Although many such solutions are expected to exist in other symmetry classes, it is not possible to obtain the global dynamics of the Hamiltonian system from direct integration. We shall show indirectly that a heteroclinic connection between steady hexagons arises as a limit of oscillating triangles if b 2 0 and c 4 + 2c 7 < 0. Let us consider solutions with triangular symmetry, i.e., v 1 = v 2 = v 3. In this case, our Hamiltonian system reduces to: v v + b 2 v 2 + (c 4 + 2c 7 ) v 2 v = 0. (42) If we choose the plus sign, there is a family of oscillating triangles bifurcating from the origin. By the global Hopf bifurcation theorem [17,23], this family of periodic solutions must either grow unbounded, reach a bifurcation point from another steady solution, or else the period must tend to infinity. Since oscillating triangles have zero average,

Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 189 bifurcation from a steady solution other than the origin is impossible. Next, we show that the family of oscillating triangles cannot be unbounded. We rewrite v in terms of its real and imaginary parts: v = x + iy. We also introduce the momentum variables p =ẋ and q =ẏ. The equation of motion is then ẋ = H/ p, ẏ = H/ q, ṗ = H/ x, q = H/ y, and the Hamiltonian is H = 1 2 (p2 + q 2 + x 2 + y 2 ) + b 2 ( 1 3 x3 xy 2 ) + 1 4 (c 4 + 2c 7 )(x 2 + y 2 ) 2. Consider now the potential part of the energy We have U = 1 2 (x2 + y 2 ) + b 2 ( 1 3 x3 xy 2 ) + 1 4 (c 4 + 2c 7 )(x 2 + y 2 ) 2. U = U x p + U y q, Ü = U x ṗ + U y q + 2 U x 2 p2 + 2 2 U x y pq + 2 U y 2 q2. Using the equation of motion, we find Ü = ṗ 2 q 2 + 2 U x 2 p2 + 2 2 U x y pq + 2 U y 2 q2. (43) If c 4 + 2c 7 < 0, then U is negative and concave for large x 2 + y 2. This means that the expression in Eq. (43) is also negative if x 2 + y 2 is large. However, at a minimum of U along a periodic solution, Ü must be positive. This means that, along every periodic solution, we obtain a priori bound for x 2 + y 2 at the point where U has its minimum, and consequently, everywhere. The a priori bounds for x and y yield a priori bounds for ẍ and ÿ by means of the equation of motion, and a priori bounds for the first derivatives follow by the identity ẋ(t) = x(t + 1) x(t) t+1 τ t t ẍ(s)ds dτ. Consequently, the family of oscillating triangles cannot be unbounded. Therefore, there must be a family of oscillating triangle solutions for which the period tends to infinity. We shall now show that this necessarily implies a heteroclinic connection between hexagons. For this, we set v = r exp(iφ) in Eq. (42), and we obtain r r φ 2 + r + b 2 r 2 cos(3φ) + (c 4 + 2c 7 )r 3 = 0, r φ + 2ṙ φ b 2 r 2 sin(3φ) = 0. (44) For oscillating triangles close to the bifurcation from the origin, φ is clearly an increasing function of t. We claim that there is a global family of oscillating triangles such that (a) r is never zero and φ is increasing monotonically with time. (b) v(t) winds around the origin once during one period. We note that as long as (a) holds, a winding number is well-defined. Since this winding number is a topological invariant, it remains constant along a continuous family of solutions. Now suppose that (a) fails at some point. Then either v(t) must cross the origin or φ must become zero at some point. If v(t) crosses the origin, it must do so at least three times in three different directions, due to the spatio-temporal symmetry of oscillating triangles

190 Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 Table 5 List of coefficients b i, b i,c i, c i of Eqs. (11) and (12) in (a) the Anderinert/silicone oil system, (b) Fluorinert/silicone oil system, (c) those of [17]. The values of l 1 and coefficients of the last column are given to two decimal places. The coefficients of the last column is given to four places in order to distinguish c 8 and c 10 (a) (b) (c) l 1 0.48 0.51 0.423 0.52 0.48303 b 2 14.613 5.24 5.79 4.11 0.76 b 3 1.079 0.34 1.26 0.97 0.081 c 3 1.489 1.17 0.86 3.1 0.0036 c 4 11.584 8.85 0.78 5.4 0.023 c 5 0.738 1.29 0.004 2.4 0.0017 c 7 12.855 9.79 2.96 4.95 0.027 c 8 0.706 0.02 0.62 0.82 0.0025 c 9 0.425 1.52 1.97 2.58 0.0033 c 10 0.715 1.48 0.06 2.63 0.0023 b 2 1.000 0.08 0.42 0.36 0.082 c 4 3.010 1.57 1.59 0.21 0.0010 c 7 3.736 2.00 1.94 0.59 0.0015 (v(t +T/3) = v(t) exp(2πi/3)). Such a solution cannot be a limit of simple loops. If φ becomes zero at some point, while φ 0 elsewhere, then necessarily φ = 0 at the same point. From Eq. (44), we infer that then necessarily sin(3φ) = 0 and φ is a constant independent of time. The orbit would therefore contain a straight line segment, which is also impossible. The only possibility left is that the period tends to infinity because the solution comes close to a steady solution. Since the only steady solutions with triangular symmetry are hexagons, this implies the existence of heteroclinic orbits between hexagons as claimed. 6. Numerical results Table 1 gives the physical parameters for each of the three systems we examine: Anderinert/silicone oil, Fluorinert/silicone oil and Fig. 2 of [4]. Tables 2 4 give further numerical data on the Takens Bogdanov points. The corresponding coefficients b i, b i,c i, c i of Eqs. (11) and (12) are listed in Table 5 We note that for the systems a and c we have listed values for two different Takens Bogdanov points; these two points are on opposite ends of the window in parameter space where Hopf bifurcations occur. That is, for values of l 1 in between these two extremes, complex conjugate eigenvalues occur, while the eigenvalues are real for values of l 1 outside this interval. The analysis of Sections 5.2, 5.3, 5.4 and 5.5 concerns case B of Section 5.1.2, i.e., the case where the coefficients of the quadratic terms vanish. In the actual physical systems, the coefficients of the quadratic terms have some finite value so that for extremely small solutions, the cubics are small compared with quadratics and case A is appropriate. For this case, numerical calculations revealed no stable patterns of very small amplitudes. For more moderate amplitudes that may be physically realizable, numerical simulations are based directly on Eqs. (11) and (12), together with those coefficients which enter into case B and are tabulated in Table 5. Although we cannot justify this procedure rigorously, we expect it to have at least qualitative validity for our situation. We note that the analysis of case B assumed that the terms involving b 2,c 4 and c 7 are the dominant ones, and indeed the table shows numerical values of these particular coefficients which are large relative to the others. For the numerical simulations reported below, we integrated the differential equations with a fourth-order Runge Kutta scheme. We chose ɛ 1 = 0.1, so that the trivial solution is always unstable, and we varied ɛ 2.

Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 191 Fig. 9. (a) Temporal evolution of v 1,v 2 and v 3 for the data of Case (a), column 1, Table 5 and ɛ 2 = 1.5. The real and imaginary parts of v i are shown. (b) Spatial evolution for the data of Case (a), column 1, Table 5, and ɛ 2 = 1.5. The patterns are in the (x y) plane. Here, and in the subsequent plots, positive contour lines are solid, and negative contour lines are dashed. The sequence in all plots is spaced in time intervals of 0.2, left to right. The initial condition was chosen to be small and without any particular symmetry; specifically, we chose the initial data v 1 = 0.02, v 2 = 0.04, v 3 = 0.01 + 0.01i and w 1 = w 2 = w 3 = 0. We then observed the long time behavior of the numerical solution which evolves from these initial data. For all the data of Table 5, we see a general trend from steady to chaotic to quasiperiodic to periodic solutions as ɛ 2 is decreased. Specifically, we found the following behaviors: Case (a), column 1: Summary. For ɛ 2 1.5, solutions become infinite in finite time. For 1.4 ɛ 2 1.4, steady hexagons are observed. For 1.5 ɛ 2 4.4, we observe chaotic, then quasiperiodic solutions, which we shall describe in more detail below. At ɛ 2 = 4.5, we find a periodic solution; in the terminology of [15], this solution is a traveling patchwork quilt (1). For ɛ 4.6, we find traveling rolls. Specifics. We now show more details of the observed patterns in the chaotic and quasiperiodic regimes. In Figs. 9 18, we show the temporal and spatial evolution of patterns for ɛ 2 = 1.5, 1.8, 2.1, 2.7, 3.2, 3.5, 3.8, 4.0, 4.4 and 4.5. In all cases, the first figure shows the temporal evolution of the amplitudes v 1,v 2 and v 3 in the complex plane as a function of time. The second figure shows a sequence of the corresponding spatial patterns; the plots are the level curves of v 1 exp(ix) + v 2 exp( ix/2 + i 3y/2) + v 3 exp( ix/2 i 3y/2). The sequence of pictures, left to right, is spaced in time intervals of 0.2. The spatio-temporal evolution for this range of ɛ 2 is complicated, but some general trends can be noted. The temporal evolution is very irregular at ɛ 2 = 1.5, and as ɛ 2 decreases, it gradually becomes more regular. At the same time, spatial patterns change from a regime dominated by triangles to one dominated by rolls. For ɛ 2 in the range from 2.1to 3.5 (Figs. 11 14), the temporal pattern appears to be a chaotic perturbation of a quasiperiodic solution. We also note that the temporal evolution of two of the components v i is very similar (e.g., v 1 and v 3 in Fig. 11 (a)), while the third one is completely different.

192 Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 Fig. 10. (a) Temporal evolution of v 1,v 2 and v 3 for the data of Case (a), column 1, Table 5, and ɛ 2 = 1.8. (b) Spatial evolution for the data of Case (a), column 1, Table 5, and ɛ 2 = 1.8. Fig. 11. (a) Temporal evolution of v 1,v 2 and v 3 for the data of Case (a), column 1, Table 5, and ɛ 2 = 2.1. (b) Spatial evolution for the data of Case (a), column 1, Table 5, and ɛ 2 = 2.1.

Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 193 Fig. 12. (a) Temporal evolution of v 1,v 2 and v 3 for the data of Case (a), column 1, Table 5, and ɛ 2 = 2.7. (b) Spatial evolution for the data of Case (a), column 1, Table 5, and ɛ 2 = 2.7. Fig. 13. (a) Temporal evolution of v 1,v 2 and v 3 for the data of Case (a), column 1, Table 5, and ɛ 2 = 3.2. (b) Spatial evolution for the data of Case (a), column 1, Table 5, and ɛ 2 = 3.2.

194 Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 Fig. 14. (a) Temporal evolution of v 1,v 2 and v 3 for the data of Case (a), column 1, Table 5, and ɛ 2 = 3.5. (b) Spatial evolution for the data of Case (a), column 1, Table 5, and ɛ 2 = 3.5. Fig. 15. (a) Temporal evolution of v 1,v 2 and v 3 for the data of Case (a), column 1, Table 5, and ɛ 2 = 3.8. (b) Spatial evolution for the data of Case (a), column 1, Table 5, and ɛ 2 = 3.8.

Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 195 Fig. 16. (a) Temporal evolution of v 1,v 2 and v 3 for the data of Case (a), column 1, Table 5, and ɛ 2 = 4.0. (b) Spatial evolution for the data of Case (a), column 1, Table 5, and ɛ 2 = 4.0. Fig. 17. (a) Temporal evolution of v 1,v 2 and v 3 for the data of Case (a), column 1, Table 5, and ɛ 2 = 4.4. (b) Spatial evolution for the data of Case (a), column 1, Table 5, and ɛ 2 = 4.4.

196 Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 Fig. 18. (a) Temporal evolution of v 1,v 2 and v 3 for the data of Case (a), column 1, Table 5, and ɛ 2 = 4.5. (b) Spatial evolution for the data of Case (a), column 1, Table 5, and ɛ 2 = 4.5. For ɛ 2 3.8 (Figs. 15 17), the motion becomes quasiperiodic. It changes to a periodic motion at ɛ = 4.5 (Fig. 18(a)). This periodic solution is a traveling patchwork quilt (1), in the terminology of [15]. The fact that such a solution is observed is somewhat remarkable, because the analysis of the Hopf bifurcation in [15] shows that the traveling patchwork quilt (1) is never stable near the bifurcation from the rest state. Of course, in the current situation, we do not see the traveling patchwork quilt (1) near a bifurcation from the rest state, but near a bifurcation from traveling rolls as discussed in the analysis of the preceding section. Case (a), column 2: Summary. Here we find steady rolls for ɛ 2 5.7, steady hexagons for 5.6 ɛ 2 3.5, chaotic solutions for 3.6 ɛ 2 3.7, and wavy rolls (1) for ɛ 2 3.8. Specifics. Figs. 19(a,b) show the spatio-temporal evolution for ɛ 2 = 3.6, and Figs. 20(a,b) show ɛ 2 = 3.8. The solution for ɛ 2 = 3.6 is temporally chaotic, while the spatial pattern has the appearance of wavy rolls (1). The solution for ɛ 2 = 3.8 is the temporally periodic wavy rolls (1). Case (b): Summary. This is the only case where no steady regime was found. We have blow-up in finite time for ɛ 2 1.8, and a chaotic regime for 1.9 ɛ 2 2.3. For 2.4 ɛ 2 2.6, we find solutions which are temporally chaotic but have a triangular spatial symmetry. For 2.7 ɛ 2 3 and again for ɛ 2 4, we have oscillating triangles. In the intervening interval, we have a traveling patchwork quilt (1) for 3.1 ɛ 2 3.9. Specifics. Figs. 21 24 show patterns found at ɛ 2 = 2.1, 2.5, 2.7 and 3.2. At ɛ 2 = 2.1, we have a temporally chaotic solution, with an approximate but not exact triangular symmetry in the spatial pattern. At ɛ 2 = 2.5, the temporal evolution is still chaotic, but the spatial pattern now has triangular symmetry and looks very much like oscillating triangles. At ɛ 2 = 2.7, we find the temporally periodic oscillating triangles solution. We note the distinct triangular shape of the periodic orbits in Fig. 23(b). This, and the appearance of chaotic solutions

Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 197 Fig. 19. (a) Temporal evolution of v 1,v 2 and v 3 for the data of Case (a), column 2, Table 5, and ɛ 2 = 3.6. (b) Spatial evolution for the data of Case (a), column 2, Table 5 and ɛ 2 = 3.6. Fig. 20. (a) Temporal evolution of v 1,v 2 and v 3 for the data of Case (a), column 2, Table 5, and ɛ 2 = 3.8. (b) Spatial evolution for the data of Case (a), column 2, Table 5, and ɛ 2 = 3.8.

198 Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 Fig. 21. (a) Temporal evolution of v 1,v 2 and v 3 for Fluorinert/silicone oil, Case (b) of Table 5, and ɛ 2 = 2.1. (b) Spatial evolution for the data of Fluorinert/silicone oil, Case (b), Table 5, and ɛ 2 = 2.1. Fig. 22. (a) Temporal evolution of v 1,v 2 and v 3 for the data of Fluorinert/silicone oil, Case (b) of Table 5, and ɛ 2 = 2.5. (b) Spatial evolution for Fluorinert/silicone oil, Case (b) of Table 5, and ɛ 2 = 2.5.

Y.Y. Renardy et al. / Physica D 129 (1999) 171 202 199 Fig. 23. (a) Temporal evolution of v 1,v 2 and v 3 for Fluorinert/silicone oil, Case (b) of Table 5, and ɛ 2 = 2.7. (b) Spatial evolution for Fluorinert/silicone oil, Case (b) of Table 5, and ɛ 2 = 2.7. Fig. 24. (a) Temporal evolution of v 1,v 2 and v 3 for Fluorinert/silicone oil, Case (b) of Table 5, and ɛ 2 = 3.2. (b) Spatial evolution for Fluorinert/silicone oil, Case (b) of Table 5, and ɛ 2 = 3.2.